Re: [patch 04/16] genirq: Introduce irq_chip.irq_compose_msi_msg() to support stacked irqchip

["Yun Wu (Abel)" <wuyun.wu@huawei.com>]

On 2014/11/18 21:25, Jiang Liu wrote:

> On 2014/11/18 21:16, Yun Wu (Abel) wrote:
>> On 2014/11/18 20:43, Jiang Liu wrote:
>>
>>> On 2014/11/18 19:47, Yun Wu (Abel) wrote:
>>>> On 2014/11/18 18:02, Thomas Gleixner wrote:
>>>>
>>>>> On Tue, 18 Nov 2014, Yun Wu (Abel) wrote:
>>>>>> On 2014/11/12 21:42, Thomas Gleixner wrote:
>>>>>>> +int irq_chip_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
>>>>>>> +{
>>>>>>> +	struct irq_data *pos = NULL;
>>>>>>> +
>>>>>>> +#ifdef	CONFIG_IRQ_DOMAIN_HIERARCHY
>>>>>>> +	for (; data; data = data->parent_data)
>>>>>>> +#endif
>>>>>>> +		if (data->chip && data->chip->irq_compose_msi_msg)
>>>>>>> +			pos = data;
>>>>>>> +	if (!pos)
>>>>>>> +		return -ENOSYS;
>>>>>>> +
>>>>>>> +	pos->chip->irq_compose_msi_msg(pos, msg);
>>>>>>> +
>>>>>>> +	return 0;
>>>>>>> +}
>>>>>>
>>>>>> Adding message composing routine to struct irq_chip is OK to me, and it should
>>>>>> be because it is interrupt controllers' duty to compose messages (so that they
>>>>>> can parse the messages correctly without any pre-defined rules that endpoint
>>>>>> devices absolutely need not to know).
>>>>>> However a problem comes out when deciding which parameters should be passed to
>>>>>> this routine. A message can associate with multiple interrupts, which makes me
>>>>>> think composing messages for each interrupt is not that appropriate. And we
>>>>>> can take a look at the new routine irq_chip_compose_msi_msg(). It is called by
>>>>>> msi_domain_activate() which will be called by irq_domain_activate_irq() in
>>>>>> irq_startup() for each interrupt descriptor, result in composing a message for
>>>>>> each interrupt, right? (Unless requiring a judge on the parameter @data when
>>>>>> implementing the irq_compose_msi_msg() callback that only compose message for
>>>>>> the first entry of that message. But I really don't like that...)
>>>>>
>>>>> No, that's not correct. You are looking at some random stale version
>>>>> of this. The current state of affairs is in 
>>>>>
>>>>>    git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git irq/irqdomain
>>>>>
>>>>> See also https://lkml.org/lkml/2014/11/17/764
>>>>>
>>>>> In activate we write the message, which is the right point to do so.
>>>>>
>>>>
>>>> I checked the current state, it seems to be the same.
>>>> Yes, the decision of postponing the actual hardware programming to the point
>>>> where the interrupt actually gets used is right, but here above I was talking
>>>> another thing.
>>>> As I mentioned, a message can associate with multiple interrupts. Enabling
>>>> any of them will call irq_startup(). So if we don't want to compose or write
>>>> messages repeatedly, we'd better require performing some checks before
>>>> activating the interrupts.
>>> Hi Yun,
>>> 	Seems you are talking about the case of multiple MSI support.
>>> Yes, we have special treatment for multiple MSI, which only writes PCI
>>> MSI registers when starting up the first MSI interrupt.
>>> void pci_msi_domain_write_msg(struct irq_data *irq_data, struct msi_msg
>>> *msg)
>>> {
>>>         struct msi_desc *desc = irq_data->msi_desc;
>>>
>>>         /*
>>>          * For MSI-X desc->irq is always equal to irq_data->irq. For
>>>          * MSI only the first interrupt of MULTI MSI passes the test.
>>>          */
>>>         if (desc->irq == irq_data->irq)
>>>                 __pci_write_msi_msg(desc, msg);
>>> }
>>
>>
>> Yes, I picked the case of multiple MSI support.
>> The check should also be performed when composing messages. That's why
>> I don't like its parameters. The @data only indicates one interrupt,
>> while I prefer doing compose/write in the unit of message descriptor.
> Hi Yun,
> 	The common abstraction is that every message interrupt could be
> controlled independently, so have compose_msi_msg()/write_msi_msg() per
> interrupt. MSI is abstracted as an special message signaled interrupt
> with hardware limitation where multiple interrupts sharing the same
> hardware registers. So we filter in pci_msi_domain_write_msg(). On the
> other handle, the generic MSI framework caches msi_msg in msi_desc,
> so we don't filter compose_msi_msg().
> 

It's true that every message interrupt could be controlled independently,
I mean, by enable/disable/mask/unmask. But the message data & address are
shared among the interrupts of that message.
Despite the detailed hardware implementation, MSI and MSI-X are the same
thing in software view, that is a message related with several consecutive
interrupts. And the core MSI infrastructure you want to build should not
be based on any hardware assumptions.

Thanks,
	Abel